Browse Items (16481 total)
Sort by:
-
Integrating LLMs into Smart Home Architecture: Design, Implementation, and Experimental Insights
Smart home systems are a key application of the Internet of Things (IoT) paradigm. These devices are either directly or indirectly connected to a network in order to perceive actions based on user willingness or sensing. As Information and Communication Technology (ICT) advances, Large Language Models (LLMs) are growing increasingly potent. Agents with LLM capabilities could be very supportive for smart home systems because of their natural language comprehension. In this paper, we introduce an LLM-powered smart home agent with real-time connectivity to smart home devices. This proposal has been experimented with and evaluated with typical smart home use cases. This experimental architecture illustrates a realtime scenario of a standard smart home, where sensors, lights, and switches are interconnected in a multi-tiered environment. The sensor, gateway, and switch modules are connected to the smart home edge server via a Web of Things (WoT) interface. The experiment has been conducted with various types of smart home use case prompts, including status requests, control requests, automation requests, and reasoning requests. The outcome of the experiment indicates that the addition of LLM to smart homes excels in natural conversational patterns compared to keywordbased agents. The prompt response time, which is unsuitable for time-sensitive tasks like anomaly detection, is a drawback, and edge LLMs could be a solution. 2025 IEEE. -
Low-Profile Metasurface-Integrated Ultra-Wideband Antenna with Enhanced Gain
In this paper a metamaterial-integrated compact antenna is proposed, and the design, simulation, and implementation are presented which works in Ultra-Wideband frequency (UWB) range. The FR4 substrate has been used to design a compact, flexible, wearable antenna. Metamaterial structure comprises of periodic arrangement of unit cells termed as metasurface, to achieve higher gain. The proposed integrated antenna exhibits maximum gain of 8.1 dBi with overall dimensions of 50 mm 40 mm. Also, the gain enhancement of 3.2 dBi along with 0.2 GHz increment in bandwidth is observed after adding the metamaterial array. Thus, the proposed antenna is suitable for wearable applications. 2025 IEEE. -
Integrating Explainable Machine Learning (XAI) in Stroke Medicine: Opportunities and Challenges for Early Diagnosis and Prevention
Stroke is a leading cause of mortality and disability worldwide, emphasizing the critical need for early diagnosis and prevention. Machine learning (ML) has demonstrated significant potential in improving stroke prediction and management by analysing complex datasets for risk stratification, diagnosis, and treatment planning. However, the adoption of ML in stroke medicine is limited by the opacity of these models, which can hinder clinical trust and decision-making. Explainable Artificial Intelligence (XAI) addresses this challenge by making ML models more interpretable and transparent, enabling healthcare professionals to understand, validate, and trust their outputs. This research work explores the integration of XAI in stroke medicine, highlighting its potential to enhance early diagnosis, personalized prevention strategies, and treatment planning. We discuss the opportunities XAI provides in identifying high-risk patients, uncovering critical predictors, and enabling informed clinical decisions. Furthermore, we examine challenges such as ensuring model reliability, addressing biases in stroke datasets, and navigating ethical considerations related to patient data privacy and algorithmic accountability. 2025 IEEE. -
Cross Domain Lexicon Transfer -A Case Specific to Application in Banking Domain
This study demonstrates the development of the financial domain lexicon and the implementation of the same in the banking sector. The study compared the working of Financial Nae Bayes Lexicon (FNB Lex) developed with 533 quarterly Earnings Call Transcript (ECT) of 16 software companies, with standard available dictionaries like VADER and Loughran-McDonald. The study showed VADERs poor discriminatory power and Loughran-McDonald with satisfactory performance. FNB Lex lexicon performed better and provided better lift over VADER and Loughran-McDonald with improved precision, recall and F1-score. 2025 IEEE. -
Self-Organizing Micro Service Composition for IoT Ecosystem
The Internet of Things (IoT) has become the central focus in many computing applications, with smart devices seamlessly integrated to meet user needs by providing services that reflect their functionalities. Service composition, the process of integrating multiple services to deliver unified functionality, is crucial in this context. However, traditional service composition techniques fall short in highly dynamic and open environments such as the IoT ecosystem, necessitating decentralized models that can effectively support service composition in such settings. The self-organizing microservice composition model for IoT addresses this need by leveraging decentralized, localized interactions that utilize bio-inspired mechanisms. These mechanisms enable the system to autonomously form complex service compositions with minimal human intervention through emergent behaviour, enhancing the systems flexibility, adaptability, and overall performance. This paper presents a model specifically designed for the IoT ecosystem, focusing on healthcare applications. The model dynamically responds to changing conditions, such as varying patient needs, device availability, and network conditions, making it highly suitable for critical healthcare environments. By providing a robust framework for managing the complexities inherent in healthcare IoT, this model has the potential to revolutionize the delivery and management of healthcare services. 2025 IEEE. -
Collaborative Model for Sustainable Energy Utilization in Cloud Infrastructure
As the infrastructures of cloud computing provides paramount services to worldwide users, persistent applications are congregated using large scale data centres at the customer sides. For such wide platforms, virtualization technique has been incorporated for multiplexing the essential sources available. Due to the extensive application variations in the workloads, it is significant to handle the resource allocation methodologies of the virtual machines (VM) for assuring the Quality of Service (QoS) of cloud. On concentrating this, the paper proposed a Decentralized Energy-Aware Collaborative Model (DEACM) for effectively managing the data centres in cloud infrastructures. Initially, the optimal model for system management and power management are declared. Then, functions of workload vectors and data collection about workloads has been carried out for optimal selection of virtual machines to migrate for balancing loads efficiently. This can be further applied for Target-based VM Migration Algorithm for determining the migrating target for VM. Moreover, the algorithm involved in energy utilization with managed QoS. The developed DEACM is evaluated using CloudSim platform and the results are discussed. The results exemplify that the DEACM can balance the workload across variety of machines optimally and provide reduced energy consumption to the complete system efficiently. 2025 IEEE. -
Enhanced Image Classification using Transfer Learning with ResNet50-V2: A Case Study on Wildlife Recognition
This study explores the application of transfer learning using the ResNet50-V2 architecture for accurate classification of Arctic wildlife species, including Arctic foxes, polar bears, and walruses. Transfer learning leverages pretrained networks to enhance performance in new tasks with limited labeled data, reducing the need for extensive data collection and computational resources. In this work, we utilized a dataset of 1000 labeled images across the three species and applied ResNet50-V2, pre-trained on ImageNet, as a feature extractor. The model achieved high accuracy, with training and validation accuracies nearing 99% and 95-97%, respectively, though minor overfitting was observed. This indicates the model's strong ability to generalize across the dataset while benefiting from pre-trained weights on diverse, non-related images. Additionally it compares with models like SSD and CycleGAN, emphasizing its capability to generalize well, handle small datasets, and mitigate overfitting. We discuss model architecture, data preprocessing, and the experimental results, focusing on improvements achievable through regularization techniques to counteract overfitting. This study demonstrates the effectiveness of transfer learning for wildlife classification, providing insights into optimizing CNNs for ecological and conservation applications. 2025 IEEE. -
Electricity Demand Prediction: An Analytical Comparison of ARIMA and Artificial Neural Network
Electricity plays a dominant role globally, especially in the economies of India. Accurately projecting its consumption is crucial for energy planning. This study focuses on forecasting electricity consumption across distinct sectors using Autoregressive Integrate Moving Average (ARIMA) and Artificial Neural Network (ANN). The efficacy of the models is evaluated via various error metrics and compared, demonstrating the superior performance of the ANN model over ARIMA model. 2025 IEEE. -
Label Informativeness-Based Minority Oversampling in Graphs (LIMO)
Class imbalance is a pervasive issue in many realworld datasets, particularly in graph-structured data, where certain classes are significantly underrepresented. This imbalance can severely impact the performance of Graph Neural Networks (GNNs), leading to biased learning or over-fitting. The existing oversampling techniques often overlook the intrinsic properties of graphs, such as Label Informativeness (LI), which measures the amount of information a neighbor's label provides about a node's label. To address this, we propose Label Informativenessbased Minority Oversampling (LIMO), a novel algorithm that strategically oversamples minority class nodes by augmenting edges to maximize LI. This technique generates a balanced, synthetic graph that enhances GNN performance without significantly increasing data volume. Our theoretical analysis shows that the effectiveness of GNNs is directly proportional to label informativeness, with mutual information as a mediator. Additionally, we provide insights into how variations in the number of inter-class edges influence the LI by analyzing its derivative. Experimental results on various homophilous and heterophilous benchmark datasets demonstrate the effectiveness of LIMO in improving the performance of node classification for different imbalance ratios, with particularly significant improvements observed in heterophilous graph datasets. Our code is available at https://github.com/smlab-niser/limo. 2025 IEEE. -
A Pipeline for Speech-to-Text Summarization and Question Identification for Enhanced Chatbot Interactions
The rapid advancements in natural language processing provide strong support for the new potential application of integrating Google Speech Recognition API, BART, and BERT to create a full pipeline for speech recognition, text summarization and question answering without breaking human interaction. The research aims to develop such a holistic pipeline involves integrating the Google Speech Recognition API to perform speech-to-text, BART for text summarization, and finally BERT for question answering based on both the summary and original transcript. The system was tested under various criteria such as testing accuracy, real-time processing performance, and stress tests for scalability where the findings include an average of 60% text compression with BART, an 88% accuracy in BERT-based question answering, and scores indicating high user satisfaction (4.3/5). Real-time processing latency can be able to cater to interaction within 2-3 seconds and the capacity of the system has proven without performance loss during simultaneous users. The research done can practically find applications in areas like intelligent virtual assistants, customer service automation and e-learning applications that improve accessibility and user experience. 2025 IEEE. -
Reinforcement Learning-Driven Innovation Clusters: Strategic Planning for Sustainable Corporate Growth
This paper explores the role of reinforcement learning (RL) in optimizing innovation clusters to foster sustainable corporate growth. We go on to establish how RL allows organizations to optimize core performance metrics (innovation output, profit growth, sustainability impact and resource allocation efficiency), and show in dynamic datasets how a network of simulated strategic decisions were made in an innovation ecosystem. Moreover, it highlights the ability of RL to adapt to ever changing industries and implement long term strategic plans besides traditional strategic practices. The results demonstrate that RL-based methods contribute to unleashing innovation and profitabilising the companies, but also to more sustainable operations, bringing into proportion the growth and social responsibility. These results demonstrate RL as an implication tool with a strong future for optimizing corporate strategies that serves as an incentive for further innovation, translating into long-term viability and success. 2025 IEEE. -
Multilevel CNN Based Hybrid Framework for Adaptive Credit Card Fraud Detection
Credit card fraud presents a substantial problem to financial organizations, as fast changing fraudulent activities necessitate advanced detection techniques. Conventional machine learning methods frequently encounter challenges with adaptability and precision in imbalanced datasets. This study presents a multilevel CNN-based hybrid architecture that combines deep convolutional networks with traditional ensemble classifiers for adaptive credit card fraud detection. The platform includes an adaptive learning module that facilitates ongoing model upgrades, guaranteeing responsiveness to emerging fraud trends. The system, evaluated using a benchmark Kaggle dataset, attained an accuracy of 99.48%, precision of 98.76%, recall of 99.05%, F1-score of 98.90%, and AUC-ROC of 99.91%, outperforming established baseline models such as Logistic Regression, Random Forest, and XGBoost. The suggested system's capacity to integrate deep feature extraction with hybrid classifiers yields enhanced detection efficiency, reduced false positives, and improved generalization. This research enhances fraud detection by overcoming the constraints of static models, rendering it applicable for real-time financial applications and adaptable to emerging threats. 2025 IEEE. -
Fine-Tuned Deep Contextual BERT for Enhanced Aspect-based Sentiment Analysis: A Comparative Study on Laptop Reviews
Sentiment analysis entails the care full analysis, conduction of interpretation and conclusion of subjective texts even as an evaluation. In the business context, the companies' strategies towards growth makes use of both level of experience of consumers, market reach, social media, opinion and reputation of the brand. The different levels of performing the analysis includes the analysis at the document, phrase, and aspect levels. The sentiment which targets the polarity on some components of texts is often recognized by various Natural language processing (NLP) tasks for example aspect level sentiment analysis. This study presents the fine-tuned deep contextual BERT (FTDC BERT) aiming at improving the accuracy of sentiment polarization prediction. We look at different types of models including the LSTM based and the attention based and the BERT based models and where they performed on the laptop dataset. The fine-tuned and pre-trained BERT model exceeded all benchmarks and gave the most accurate work at 84.48%. This remarkable achievement testifies to the capability of the model in adapting its structure to varying degrees of sentiment contained in laptop reviews. Based on the comparative analysis, different models have different degree of success which indicates that sentiment has to be modelled separately for every set of data. This paper describes interesting areas of the future inline sentiment analysis for researchers and practitioners. 2025 IEEE. -
An Optimized Approach for Spam Message Detection Using C4.5 Classifier with Stochastic Hill Climbing and Genetic Algorithm for Feature Selection
In the mobile industry, text messaging is a popular feature that is mainly intended to make money for service providers. But spam, which is defined as unsolicited bulk messages that contain commercial content, has become a widespread problem. These spam texts are frequently used to spread phishing links or advertise goods and services in order to make money. The phone alerts the user whenever spam text messages arrive in their inbox. When the user discovers that the message is unsolicited, these unsolicited texts not only take up storage space and waste their time, but they also irritate them. Even with the development of numerous sophisticated algorithms to identify spam, users are still impacted by text message spam. Thus, the mobile sector needs to implement efficient filtering methods. The proposed study uses the C4.5 Decision Tree as the classification model and combines a Genetic Algorithm and Stochastic Hill Climbing to find optimal features in order to detect spam in text messages. This method uses metaheuristic techniques to find the best features, which are then categorized using decision trees. This hybrid model performs better than current classification methods. 2026 IEEE. -
Robust Rice Leaf Disease Detection using Advanced Preprocessing and Deep CNNs for Class Imbalance Resolution
This study addresses the growing challenges posed by plant diseases, particularly in the rice industry, which is vital for many communities. The research propose a robust framework that integrates Deep Convolutional Neural Networks (Deep CNN) with advanced preprocessing techniques to identify rice leaf diseases, including Brown Spot, Leaf Blast, Hispa, and healthy leaves. Our approach employs normalization to enhance convergence during training and data augmentation to improve model generalizability. Additionally, implement the Synthetic Minority Over-sampling Technique (SMOTE) to create synthetic samples for under-represented classes, addressing class imbalance within the dataset. Experimental results demonstrate the model's impressive accuracy, achieving 98.2% for Brown Spot, 97.5% for Leaf Blast, 94.3% for Hispa, and 96.8% for healthy leaves. Furthermore, our method outperforms established CNN architectures such as AlexNet, VGG16, and ResNet50, showcasing the effectiveness of sophisticated preprocessing in enhancing plant disease detection systems and supporting food security initiatives. 2025 IEEE. -
Data-Driven Insights into Student Performance: Benchmarking Machine Learning Models for Grade Prediction using Regression and Classification Approaches
This research explores the effectiveness of 17 machine learning models in predicting student performance across Mathematics and Portuguese datasets. The primary goal of this study was to evaluate and compare regression and classification models to identify the most accurate predictors of student grades. A range of algorithms was tested, including linear models (Linear Regression, Elastic Net, Ridge, Lasso), tree-based models (Random Forest, Gradient Boosting, CatBoost, LightGBM), and advanced techniques (Neural Networks, SVM, XGBoost, Naive Bayes, SVR). The methodology involved data preprocessing, feature engineering, and splitting data into training and test sets. Base models were implemented, followed by hyperparameter tuning to optimize performance. Metrics like RMSE, MAE, MSE, R2 (for regression), and accuracy, precision, recall, F1 score (for classification) were used to assess performance. The study found that Gradient Boosting and Elastic Net consistently outperformed other models in regression tasks, achieving the highest R2 scores. For classification, Logistic Regression proved to be the most accurate, followed by Naive Bayes. These findings provide valuable insights for model selection in educational performance prediction, establishing Gradient Boosting and Logistic Regression as benchmark models. 2025 IEEE. -
Predictive Analytics for Stock Price Forecasting: Machine Learning Techniques in Financial Markets
Stock market forecasting is significantly challenging because financial data generally exhibits non-linearity, and volatility is highly presented. Traditional methods such as the ARIMA model and NN fail to take a good grasp of intricate and complex temporal patterns in changes related to market trends. By overcoming these limitations, it makes use of LSTM and combines GAN networks. The LSTM exploits the historical stock price data for temporal dependencies, whereas GAN produces realistic synthetic data to augment model training. The Stock Market Dataset was used, and the proposed model was implemented using Python with TensorFlow and PyTorch frameworks. The hybrid LSTM-GAN model resulted in better performance with an RMSE of 0.0125, MAE of 0.0093, and R2 of 0.926, thus outperforming LSTM and traditional forecasting models. This work greatly enhances the accuracy of forecasting, avoids overfitting, and promotes performance in volatile market environments. The results are extremely useful for investors, financial analysts, and trading platforms because they can make better predictions. 2025 IEEE. -
An Intelligent Model for Detecting Cervical Cancer Using U-Net Segmentation and YOLO Classification Augmented with Lion Optimization
One of the leading causes of death for women is cervical cancer, and survival rates are significantly increased by early identification using Pap smear analysis. However, manually reviewing Pap smear images takes time and is prone to mistakes. Using an enhanced Convolutional Neural Network (CNN) with Transformer-like classification and U-Net-based nucleus segmen-tation, this study suggests an automated cervical cancer detection system. To increase feature extraction, the image is preprocessed using techniques such as edge detection, contrast enhancement with CLAHE, and greyscale conversion. The PR-processed image is segmented using U Net segmentation. A YoLo-based CNN optimised with the Lion optimiser (Evolved Sign Momentum) is used to classify the segmented nuclei to improve convergence and accuracy. Results from experiments show that our model outperforms con-ventional methods in terms of classification accuracy. By lowering reliance on manual screening and increasing early diagnosis rates, this automated approach can help medical practitioners detect cervical cancer more quickly and accurately. 2025 IEEE. -
Enhancing Stock Price Prediction: A Multi-Model Framework Integrating Technical Indicators and Sentiment Analysis
This paper proposes a multi-model strategy that would improve the predictive power of stock prices by combining time-series analytics with external market indicators. The system allows five different base prediction methods; Long Short-Term Memory (LSTM), Enhanced Bidirectional LSTM (XLSTM), Support Vector Machine (SVM) which may use radial basis function (rbf), linear or polynomial (poly) kernels, Autoregressive integrated moving average (ARIMA), and Seasonal Autoregressive integrated moving average (SARIMA). A stacking procedure which uses linear regression as a meta-model together with a voting ensemble method is then employed to link these base models. The feature engineering is thorough, as it provides for general price and volume data, a battery of technical indicators (SMA10, SMA20, EMA 12, EMA 26, MACD elements, and RSI14) and a general sentiment indicator (summarised financial news). Sentiment analysis is performed by a pipeline that is trained using RoBERTa and yields discrete numerical values (0 negative, 1 neutral, 2 positive). The model's capability is very rigorously gauged by the conventional metrics Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Directional Accuracy (DA). The real-world results demonstrate that the ensemble method is very efficient where the stacking arrangement leads to the lowest total MAPE of 0.6027 % MSFT and the highest directional Accuracy of 75.86 % GOOGL, thus, providing a strong evidence for the effectiveness of the thorough integration of heterogeneous machine-learning, statistical, and sentiment- analysis methods to produce the most accurate financial forecasts. 2026 IEEE. -
A Systematic Approach for Enhancing the Curriculum Development based on the Gap Analysis to Meets the Standards of Accreditation
The article focuses on the identification of gaps in course outcomes and program outcomes in an outcome-based education. The identified gaps help in the redesigning of the curriculum as per the standard of accreditation, new education policy of India and industry-academia related gap. The gaps are identified on the targets and the attained value of course outcome of a particular course. The course outcomes are in turn related to program outcomes and on this basis, we could be able to identify the program outcomes which are not attained. This helps us in the formulation of new curriculum or redesigning of existing curriculum. In this research paper we will discuss how to calculate the attainment of any particular course, attainment of program outcomes, identification of gaps and the suggest the action plan to fill this gap through the revision of curriculum. 2026 IEEE.